Release Notes
Version 4.0.1 - beta release, please report issues
1) The wellstate approach for relativistic calculations has been modified in light of the wellstate changes described below. The confining well potential and node counts found for J=L+1/2 are now carried over and used for J=L-1/2 to avoid possible inconsistencies in the corresponding projectors.
2) Several minor bugs (uninitialized variables, single-precision constants, potential array overwrites, overflowing buffers, etc.) have been corrected. Thanks to J. E. Pask and A. C. Medina
3) modcore2.f90 and modcore3.f90 have been modified to remove numerical noise in the 3rd and 4th derivatives that could influence the output on the linear grid at very small radii.
Version 4.0.0 - beta release, please report issues
1) The code has been reorganized to permit 3 or more projectors per angular momentum L. Bound states in successive shells and scattering states are processed together in loops over [1:nproj(l1)] throughout the code. This replaces the separate treatment of the first and second projectors used in earlier versions. The maximum allowed nproj is set by mxprj, presently 5, in oncvpsp.f90. (Going beyond 5 is probably useless, and would require modifying some format statements by hand.)
The following routines have been reorganized in this manner:
gnu_script.f90 run_diag_sr_so_r.f90
gnu_script_r.f90 run_ghosts.f90
linout.f90 run_optimize.f90
linout_r.f90 run_phsft.f90
oncvpsp.f90 run_phsft_r.f90
oncvpsp_nr.f90 run_plot.f90
oncvpsp_r.f90 run_plot_r.f90
psatom.f90 run_vkb.f90
psatom_r.f90 run_vkb_r.f90
run_config.f90 sbf_basis_con.f90
run_config_r.f90 sr_so_r.f90
run_diag.f90 upfout.f90
run_diag_r.f90 upfout_r.f90
2) The overall optimization approach is retained, in the sense that each successive projector pseudo wave function (ppwf) is calculated independently enforcing off-diagonal overlap constraints with lower-index (presumably lower energy) ppwfs. The underlying spherical Bessel function basis is calculated based on the first ppwf's value and slope constraints at rc as in the two-projector code, but is expanded at each stage to retain the same number of "null" functions for consistent variational flexibily in convergence optimization. Convergence is typically found to be rather uniform, improving on that of previous two-projector results.
3) The wellstate algorithm for simulating scattering states with normalizable bound states for convergence optimization has changed. A "soft step" potential is added to the all-electron potential as before, but the asymptotic potential is a fixed energy (presently 0.5 Ha) above the specified scattering-state energy. Only the width of the step is adjusted to match that energy. In the previous version, the number of nodes of this simulated bound state was fixed at one more than the next-lower ppwf. The new algorithm explores this possibility first, but then allows successively more nodes. The new algorithm is much more robust, and will converge for considerably higher energies and larger intervals. Applied to input data that were developed for the old algorithm, however, the results are very similar. At present, the input data ep is used for the first well state if there are no bound ppwfs. Otherwise the first is debl above the last bound ppwf as in older versions, and higher well states are debl above each other. If a highly-localized bound state is found when searching for the first specified scattering state, it is substituted for that state since ghosts and poor performance would otherwise invariably result.
4) More projectors are an invitation to ghost states. The ghost-detection routine from version 3.3.1 has been modified to ensure convergence when deeper core shells are treated as valence. Ghost-detection has been incorporated in the fully-relativistic code. Note that a legitimate low-energy scattering resonance may be reported as a positive-energy ghost, especially for L=3.
5) The graphics output has been modified. The usual semi-local plus local potential and charge density plots are displayed first. Then, for each L, the all-electron and pseudo wave functions are compared for successive projector energies, either bound or scattering, and labeled by energy in the title. Finally, the orthogonal projectors are displayed, with their coefficients in the title. These plots span only radii 0 to 1.1*rc, and give a better idea of the real-space scales required if projectors are to be represented accurately on a real- space mesh. The radial log mesh, which has been made denser to deal with deeper cores, is now "sampled" for these plots to reduce the total output size. The final plot is the total kinetic energy error versus plane-wave cutoff for the first projector as in previous releases. The new graphics is illustrated in 73_Ta_plots.pdf in this directory.
6) The excited-configuration tests have been significantly modified. The solution if the radial Schroedinger equation for the pseudo wave functions is more challenging with more projectors, and must be started with tight energy bounds around the expected solution. This is achieved by starting all excited-atom runs from the reference configuration, and "adiabatically" varying the occupation numbers concomitantly with self-consistency convergence. This is in fact done in two stages, initially from the reference to the most highly ionized intermediate configuration, and then from that to the final one. A more detailed description is in psatom.f90.
7) The fact that the scalar-relativistic all-electron radial Schroedinger equation contains a first-order derivative operator violates the conditions under which generalized norm conservation is exact. This is seen in the "B matrix Hermiticity error" in the *.out file, and treated by an ad-hoc symmetrization. Inner shells of heavy atoms reveal a related problem - projectors do not go to zero at rc as they do in a non-relativistic calculation. A plausible ad-hoc fix for this is to smoothly blend an effective potential into the projector calculation over the last 5% of [0:rc]. This potential is calculated in vrel.f90. These issues eventually limit the accuracy of the psp at sufficiently small rc.
8) Thanks to a contribution from A. C. Medina, the libxc capabilities are now extended to libxc versions 4.0.0 (just a coincidence) and above. To use earlier versions, make.inc must be modified as indicated. The test/data files 38_Sr_lxc.dat and 79_Au_lxc.dat require libxc.
9) The tests/refs *.out files are computed with the code in this release, and show small differences from previous releases. While the psps have not all been tested in benchmark calculations, enough testing has been done to expect very comparable performance. Several datasets are new or have been modified to demonstrate new capabilities. Input data from the two presently published oncvpsp periodic tables runs without error.
http://www.pseudo-dojo.org/
http://www.quantum-simulation.org/potentials/sg15_oncv/
10) The unscreened local potetial at the origin of the linear output mesh is now set to the value on the first point of the internal log mesh. This suggestion of F. Gygi partially mitigates the erratic behavior of GGA potentials arising from very small values of the psedo-chargre density near the origin. (Even a small model core charge completely eliminates this problem.)
11) For convenience in testing many sets of input data, tests/multirun.sh
and tests/multirun_r.sh will launch groups of NPROC=4 jobs at once with
a command like
12) I thank John Pask for stimulating my interest in this extension of oncvpsp, and for a great deal of testing and useful feedback in the course of the its development.
Version 3.3.1
1) The recently-introduced ghost detector has been upgraded to scan positive energies as well. Adjustments have also been made to overcome a few seldom-encountered errors reported to me. See doc/ghosts.txt for details, including advice on fixing ghosts found using published oncvpsp input data.
2) The log-derivative routines have been slightly modified since ghosts found by the detector would occasionally fail to show up as steps in the pseudopotential log-derivative plots. This primarily involved reducing the radii at which these derivatives are compared, so the plots will differ slightly from those of previous releases. The energy ranges have been increased for some of the tests/data examples.
3) An indexing bug in the diagnostic tests which could change the test radii from those intended has been fixed. Some numbers in the Diagnosics will change as a result.
4) vkboutwf was modified to remedy a node-counting problem at very small core redii (eg., treating 1s as valence for a first-row atom)
5) Perdew-Wang LDA using libxc with iexc==-001012 is now accepted for upf output for pwscf.
6) Errors that cause stops will now write "ERROR" with some identification on either the standard output or standard error (both go to the *.out file in my scripts). This will facilitate examination of results in which many sets of data are run by automated scripts with a simple "grep" command. WARNINGs remain in caps as well.
7) The optimize to old_optimize comparison has been deactivated in run_optimize.f90, given extensive experience with the corrected routine as discussed under 3.3.0 below. I will continue to carry old_optimize.f90 for a while and it can be simply substituted by moving the comment "!"s following lines 171 and 346 in run_optimize. This will reproduce the results of 3.2.3 while including the new ghost diagnostics and other improvements above.
8) While the tests/refs/*.out files differ from 3.3.0 reflecting these changes, the embedded pseudopotential output files have no significant differences.
Thanks to J. E. Pask and M. van Setten for their contributions.
Version 3.3.0
1) An unfortunate math error made its way from my notes into the oncvpsp paper and into optimize.f90. Fortunately, the only effect of this has been to keep this routine from reaching the exact minimum of the residual kinetic energy above the cutoff qc, which has proven a very good proxy for plane-wave code convergence. The corrected routine also significantly simplifies the algorithm to find the minimum. Details are explained in Erratum_PhysRevB.88.085117.pdf in this directory which has been submitted. The new version of run_optimize.f90 calls the renamed old_optimize.f90 and reports the new minimum and the typically extremely small difference. Benchmark tests against all-electron results using input designed with the old code show insignificant differences. The 3.2.3 results can be reproduced by commenting out lines 177-180 and 293-296 of run_optimize. The preceding old_optimize calls will be dropped in some later release.
2) Deep-lying ghosts far below the energy range of the usual log-derivative plots have surprised a few users with bizarre results in applications. A new ghost detector has been designed which should reliably detect these. It is described in ghosts.txt in this directory.
Version 3.2.3
1) An error common to modcore2 and modcore3 has been corrected. The intended "crossover blend" from the Teter function to the all-electron core charge tail was incorrectly coded. The result was that the blend occurred far out in the tails of both functions, so that for all practical purposes, the model core was "pure Teter." Correcting this to the intended blending region (from the model-core / AE-core crossing radius to the first zero of the Teter function) produced significant undesirable structure in the derivatives of the blended model charge, likely to cause convergence errors, and especially acoustic sum rule violations in phonon calculations. The new versions of these routines simply use the Teter function for all radii, which essentially reproduces the results of 3.2.2 and 3.2.1.
2) The core radii in the input data echo in .upf files have been corrected. Previous versions erroneously reported rcs as shifted to the next-larger log mesh point for internal use, which could create inconsistencies due to rounding when used to reproduce these potentials. The original input rcs are now truncated to 5 decimals before use, and reported to 5 decimals in the .out file and the *.upf echo, ensuring consistency in re-use.
3) A similar echo of the input data is now appended to the end of the psp8 files, and is ignored by abinit. (The actual log-grid-point rcs are given as "r_core=" on the first psp8 line, and the input rcs in the trailing echo.)
4) The psfile input datum may now have the values "psp8," "upf," and "both." The extract.sh script will now extract both psp8 and upf potential files when both are present in the *.out file. Two tests/data examples have been changed (Ge and Hg) to use this option.
5) In keeping with the conventions of ABINIT version 7.10.5 and later, the extension_switch in line 6 of psp8 files is now 1 (non-relativistic) or 3 (relativistic) to indicate the presence of the valence pseudo-charge density (added in 3.1.1). This must be changed to 0 (nr) or 2 (r) for earlier ABINIT versions.
6) The valence charge listing in psp8 files is now supplemented by the all-electron valence charge and all-electron core charge in two additional columns. This is intended for auxiliary programs which do Bader and related analyses of charge densities. It is signaled by a 2nd integer following the "extension_switch" integer on line 6 of the psp8 file which is ignored by abinit.
7) Numerous small corrections to enhance inter-platform transferability have been incorporated (uninitialized variables, single-precision constants, overreaching array bounds, etc.) (Thanks to J. E. Pask)
8) A simple modification of FFLAGS in make.inc allows the use of newer libxc releases. This has been tested with libxc-2.2.1, used in abinit-8.0.8. The list of XC functions and their identifying 3-digit numbers in libxc_use.txt has been updated to 2.2.1. Be sure to uncomment the appropriate "LIBS +=" line for your make.inc. upfout.f90 has not been updated to include more functionals for pwscf (volunteers?).
9) Be aware that two sets of well-tested oncvpsp potentials have been published for most of the periodic table and are continuing to be expanded:
http://www.abinit.org/downloads/pseudodojo/pseudodojo
http://www.quantum-simulation.org/potentials/sg15_oncv/
The input data is included on these web sites, and in general will work fine to drive this release to get an alternative output format, a different Exc functional, or relativistic potentials with spin-orbit.
While I expect the sg15 potentials will be updated to reflect the error corrected in item 2 above, be sure you are using the actual input data files rather than the echo in their upf files (at least until their rcs appear in 5-decimal format).
(Thanks for suggestions and feedback to M. van Setten, M. Giantomassi, M. Verstraete and F. Gygi.)
Version 3.2.2
1) The new core features crashed for the relativistic treatment of a partiularly challenging data set one of you came up with. The solution was to use multiplicity-weighted averages of the fully-relativistic state-by-state all-electron charges and valence pseudocharges in the Teter metric analysis. There is essentially no difference in the output.
Efficient coding of this scheme required a significant reorgainzation of oncvpsp.f90 and modcore.f90, but the underlying math is unchanged except as noted above. Scalar-relativistic output from the built-in tests is identical, and fully-relativistic has very insignificant differences in the Teter analysis and hence icmod==4.
2) Minor bug fixes.
Version 3.2.1
1) The shell script tests/data/TEST.sh has been introduced. This will run all the tests with plotting suppressed, and summarize the results in TEST.report. The new build procedure (see INSTALL) automatically runs this. (Thanks to M. Verstraete)
2) Several new options and tests have been added for model core charges. These are discussed in detail in core_correction.txt in this directory. Some tests/data and tests/ref files have been modified to test these. (Thanks to M. van Setten and M. Giantomassi for motivating and testing these additions)
3) The upf output now includes pseudo wave funtions for the occupied states. This will allow LDA+U calculations in some codes. Please note that ONCVPSP potentials have not been tested or benchmarked for this use. Also be aware that LDA+U is a semi-empirical mean field theory, and does NOT represent many-body phyysics. (Thanks to R. Sundararaman for providing and testing these code modifications)
4) The ability to run totally non-relativistic calculations was inadverently disabled in 3.0.0 and has been restored. This can be used to verify that the small errors discussed following Eq.(24) in my paper arising from the scalar-relativistic (and relativistic) issues with general norm conservation are corrected.
This also applies to the accuracy of convergence of the unscreened potentials to Coulombic beyond r_c discussed in item 6 below.
5) The automated plots now include log-derivative comparisons for angular momenta from lmax+1 to 3. These are calculated using the local potential, and may be used to decide if lmax should be increased. They are placed before the convergence plot. This motivated changes in two tests/data files.
6) An inconsistency in tne unscreening of the pseudopotentials due to the slightly longer tail of the all-electron charge density has been fixed in an appropriate ad-hoc manner. This was totally inconsequential and only showed up when rlmax was set to a large value. This should never be necessary since the (corrected) pseudopotentials are very well converged to Coulombic beyond the largest r_c. This change is in oncvpsp*.f90, following the comment "fix unscreening error ..."
The underlying reason for the non-overlapping tail region arises from the fact that the bound-state radial eqution solvers begin their inward integration at a radius which is 10X that of the classical turning point. When this point is inside r_c, this will be different for the all-electron and pseudopotentials. The charge densities in this region are negligible (~1e-12) but their one-third power in the exchange-correlation potential is large enough to notice. (Thanks to C. Fortmann)
7) A typo in a format statement in oncvpsp_r.f90 which bothered some compilers has been fixed. (Thanks to G. Marco)
Version 3.1.1
1) The routine dp3int which performed cubic polynomial interpolation from the internal log radial grid to our preferred linear output grid has been upgraded to enable nth-order interpolation, with the internal parameter npoly presently set to 7. This allows the polymial model core charge for the non-linear core correction to be exact on the linear grid, and eliminates noise when the application code take derivatives internally. dpnint is also used in eresid. The changes in the output are negligible. The output routines linout, linout_r, upfout, and upfout_r have been "cleaned up" to take further advantage of dpnint. (Thanks to Matthieu Verstrate and Alberto Garcia for calling my attention to the external-derivative issue.)
2) The gnu_script* routines have been modified to take advantage of features in gnuplot-4.6 (and presumably later) and to improve the visibility of the graphics. With older versions, changing the text 'set termoption dash' to '#set termoption dash', line 137 in gnu_script.f90 and 181 in gnu_script_r.f90 should suffice to run the graphics without dashed lines. Several intermediate points have been added to the "Energy Error" convergence plots. The file docs/40_Zr_plots.pdf illustrates the new graphics.
3) The valence pseudo-charge density has been appended to the end of the psp8 output for potential future use in abinit or other codes.
Version 3.0.0
1) The major advance of this version is the contribution by Matthieu Verstraete of an interface to the libxc library of exchange-correlation functionals. Instructions for the option of installing this are in the INSTALL file. For consistency with libxc, the built-in exc.f90 routines have been redone, basically for more significant figures in the constants. excpwca was updated in this manner and renamed excpzca to reflect the fact that it actually computed the Perdew- Zunger-Ceperly-Alder lda referenced in its comments. Changes in the ref/.out files are in the 5-6th decimal place. Details are in the new libxc_use.txt file in this directory.
2) All system-dependent editing in now confined to the make.inc file
3) The scripts directory in the main directory has been introduced. This has copies of the shell scripts in tests, and fldiff.pl, borrowed from Abinit for the new compare.sh script which allows easy testing of your installation. The shell script set_path must be run in the main directory to set the correct paths in these scripts.
4) Minor bug fixes, none of which detectably influence the output have been included, again with thanks to Matthieu Verstraete.
5) Some of the coding standards described in the file in this directory have been relaxed to expidite the libxc interface (multiple routines in one file, modules, *.F file to be pre-processed), but I would prefer that this be an exception rather than a model for future contributions.
6) As you must already know, the directory into which the tarball now unpacks is tagged with the version number.
Version 2.1.2
1) A bug has been fixed so that the "First projector wave function outermost peak radius" diagnostic result is calculated correctly for positive-energy well-bound states. Nothing else in the output changes.
Version 2.1.1
1) The routine sbf8 has been replaced with an improved version that gives an overall speedup of program execution by ~X5!
The original sbf8 was written decades ago (in connection with my development of FLAPW) using a recursive algorithm for spherical Bessel functions which retains accuracy at large l values encountered in this application. For large values of the argument (not usual there) it becomes very slow. The routine eresid has dominated the run time of this whole code, calling sbf8 inside a large double loop with small l and mostly large arguments. The small-l cases (l = 1-3) are now treated with the sin-cos expressions for spherical Bessel functions, which are much faster for large arguments. The new and old versions agree to essentially machine accuracy.
2) I have been informed that abinit-7.6.3 will incorporate the changes to use psp8 files with spin-orbit, and this and subsequent releases will no longer need a patch.
3) I'm now encouraging users to submit input data for new elements.
4) The non-relativistic script run_nr.sh now creates
5) A minor change to the run*.sh and replot.sh scripts eliminated a possible problem with gnuplot.
Version 2.0.2
1) Minor bug fix in pspot: In line 125, finite small eps in if statement caused screened semi-local psps to be set to zero at very small radii.
2) Minor bug fix in run_diag: For some choices of rc and positive second-projector energies, scattering states can be nodeless inside rc. If lschvkbbe returns an error searching for a state with a node that matches the all-electron log derivative, it is called again to search for a nodeless one before run_diag reports an error.
Neither of these bugs had any effect on the pseudopotentials in the examples, but there were a few spurious diagnostic error reports.
3) A few small parameter adjustments were made to the examples to take advantage of the new convergende graphics and make the convergence more uniform in a few cases
Version 2.0.1
1) A fully relativistic version giving psps including spin-orbit effects has been implemented (executable src/oncvpspr.x, script tests/run_r.sh). The same input files drive the non- scalar- and fully-relativistic codes. upf-format psp files have been tested with quantumespresso-5.0.4 PWSCF. psp8 files for ABINIT-7.4.3 require the supplied patch be applied. Hopefully these changes can be incorporated in later abinit releases. Details are in doc/relativistic.
2) Ghost detection using the arctan(log derivative) plots has been made more robust. In particular, the occasional occurrence of false-positive with the relativistic psp8 potentials, and has no effect on the results for solids. (The OPTION local variable in run_vkb can be reset to restore the behavior of earlier releases.)
5) Minor changes in the main routines (oncvpsp*) correct their behavior when a negative-energy is specified for the second projector and this is not just an occupied valence state over a semi-core being treated as valence. A WARNING will be issued since this is probably inadvertent, and it is probably better to adjust debl for a small positive energy.
Version 1.1.1
No real changes to the code
1) Phys. Rev. B reference now printed in output header and UPF files
2) Minor cosmetic bug-fix in linout.f90
Version 1.1.0
1) Pseudopotential output in the UPF format has been implemented so that ONCVPSP psps can be used in PWSCF calculations. The input variable psfile must now be appended to the first line of the input data with values psp8 for ABINIT and upf for PWSCF (www.quantumespresso.org). the script tests/extract.sh detects which format is in the *.out file and extracts it as before. Tests comparing PWSCF and ABINIT results give excellent agreement.
2) Some related changes are made in the documentation.
Version 1.0.2
No real changes to code
1) arXiv reference now printed in output header
2) Revised preprint in doc
3) Some labeling corrected in tests/data/
Version 1.0.1
Very minor changes
1) Main program now prints the radius of outermost peak of all-electron wave function for the first projector as a guide for choosing rc for each l.
2) Inconsistencies in the comments in some of the tests/data/*.dat files have been corrected. None of the data has been changed.
3) tests/data/03_Li.dat has been added, but not tested in any calculations for solids.
3) Item 2 under "Potential Issues" in doc/users_guide.txt has been expanded.